站长资源数据库
Redis偶发连接失败案例实战记录
前言
本文主要给大家介绍了关于Redis偶发连接失败的相关内容,分享出来供大家参考学习,下面话不多说了,来一起看看详细的介绍吧
【作者】
张延俊:携程技术保障中心资深DBA,对数据库架构和疑难问题分析排查有浓厚的兴趣。
寿向晨:携程技术保障中心高级DBA,主要负责携程Redis及DB的运维工作,在自动化运维,流程化及监控排障等方面有较多的实践经验,喜欢深入分析问题,提高团队运维效率。
【问题描述】
"htmlcode">
CRedis.Client.RExceptions.ExcuteCommandException: Unable to Connect redis server: ---> CRedis.Third.Redis.RedisException: Unable to Connect redis server: 在 CRedis.Third.Redis.RedisNativeClient.CreateConnectionError() 在 CRedis.Third.Redis.RedisNativeClient.SendExpectData(Byte[][] cmdWithBinaryArgs) 在 CRedis.Client.Entities.RedisServer.<>c__DisplayClassd`1.
"color: #ff0000">【问题分析】
同时间,服务器端显示Redis服务端有丢包现象:345539 – 344683 = 856个包。
Sat Apr 7 10:41:40 CST 2018 1699 outgoing packets dropped 92 dropped because of missing route 344683 SYNs to LISTEN sockets dropped 344683 times the listen queue of a socket overflowed
Sat Apr 7 10:41:41 CST 2018 1699 outgoing packets dropped 92 dropped because of missing route 345539 SYNs to LISTEN sockets dropped 345539 times the listen queue of a socket overflowed
"color: #ff0000">【关于backlog overflow】
在BSD版本内核实现的tcp协议中,server端建连过程需要两个队列,一个是SYN queue,一个是accept queue。前者叫半开连接(或者半连接)队列,在接收到client发送的SYN时加入队列。(一种常见的网络攻击方式就是不断发送SYN但是不发送ACK从而导致server端的半开队列撑爆,server端拒绝服务。)后者叫全连接队列,server返回(SYN,ACK),在接收到client发送ACK后(此时client会认为建连已经完成,会开始发送PSH包),如果accept queue没有满,那么server从SYN queue把连接信息移到accept queue;如果此时accept queue溢出的话,server的行为要看配置。如果tcp_abort_on_overflow为0(默认),那么直接drop掉client发送的PSH包,此时client会进入重发过程,一段时间后server端重新发送SYN,ACK,重新从建连的第二步开始;如果tcp_abort_on_overflow为1,那么server端发现accept queue满之后直接发送reset。
通过wireshark搜索发现在一秒内有超过2000次对Redis Server端发起建连请求。我们尝试修改tcp backlog大小,从511调整到2048, 问题并没有得到解决。所以此类微调,并不能彻底的解决问题。
【网络包分析】
我们用wireshark来识别网络拥塞的准确时间点和原因。我们已经有了准确的报错时间点,先用editcap把超大的tcp包裁剪一下,裁成30秒间隔,并通过wireshark I/O 100ms间隔分析网络阻塞的准确时间点:
"background-color: #ffff00">"background-color: #ffff00">"background-color: #ffff00">"background-color: #ffff00">"background-color: #ffff00">"background-color: #ffff00">"background-color: #ffff00">"background-color: #ffff00">"color: #ff0000">【进一步分析】
为了了解这1.43秒之内,Redis Server在做什么事情,我们用pstack来抓取信息。Pstack本质上是gdb attach. 高频率的抓取会影响redis的吞吐。死循环0.5秒一次无脑抓,在redis-server卡死的时候抓到堆栈如下(过滤了没用的栈信息):
Thu May 31 11:29:18 CST 2018
Thread 1 (Thread 0x7ff2db6de720 (LWP 8378)):
#0 0x000000000048cec4 in "htmlcode">clientsCron(server.h): #define CLIENTS_CRON_MIN_ITERATIONS 5 void clientsCron(void) { /* Make sure to process at least numclients/server.hz of clients * per call. Since this function is called server.hz times per second * we are sure that in the worst case we process all the clients in 1 * second. */ int numclients = listLength(server.clients); int iterations = numclients/server.hz; mstime_t now = mstime(); /* Process at least a few clients while we are at it, even if we need * to process less than CLIENTS_CRON_MIN_ITERATIONS to meet our contract * of processing each client once per second. */ if (iterations < CLIENTS_CRON_MIN_ITERATIONS) iterations = (numclients < CLIENTS_CRON_MIN_ITERATIONS) "htmlcode">clientsCronResizeQueryBuffer(server.h): /* The client query buffer is an sds.c string that can end with a lot of * free space not used, this function reclaims space if needed. * * The function always returns 0 as it never terminates the client. */ int clientsCronResizeQueryBuffer(client *c) { size_t querybuf_size = sdsAllocSize(c->querybuf); time_t idletime = server.unixtime - c->lastinteraction; /* 只在以下两种情况下会Resize query buffer: * 1) Query buffer > BIG_ARG(在server.h 中定义#define PROTO_MBULK_BIG_ARG (1024*32)) 且这个Buffer的小于一段时间的客户端使用的峰值. * 2) 客户端空闲超过2s且Buffer size大于1k. */ if (((querybuf_size > PROTO_MBULK_BIG_ARG) && (querybuf_size/(c->querybuf_peak+1)) > 2) || (querybuf_size > 1024 && idletime > 2)) { /* Only resize the query buffer if it is actually wasting space. */ if (sdsavail(c->querybuf) > 1024) { c->querybuf = sdsRemoveFreeSpace(c->querybuf); } } /* Reset the peak again to capture the peak memory usage in the next * cycle. */ c->querybuf_peak = 0; return 0; }如果redisClient对象的query buffer满足条件,那么就直接resize掉。满足条件的连接分成两种,一种是真的很大的,比该客户端一段时间内使用的峰值还大;还有一种是很闲(idle>2)的,这两种都要满足一个条件,就是buffer free的部分超过1k。那么redis-server卡住的原因就是正好有那么50个很大的或者空闲的并且free size超过了1k大小连接的同时循环做了resize,由于redis都属于单线程工作的程序,所以block了client。那么解决这个问题办法就很明朗了,让resize 的频率变低或者resize的执行速度变快。
既然问题出在query buffer上,我们先看一下这个东西被修改的位置:
readQueryFromClient(networking.c): redisClient *createClient(int fd) { redisClient *c = zmalloc(sizeof(redisClient)); /* passing -1 as fd it is possible to create a non connected client. * This is useful since all the Redis commands needs to be executed * in the context of a client. When commands are executed in other * contexts (for instance a Lua script) we need a non connected client. */ if (fd != -1) { anetNonBlock(NULL,fd); anetEnableTcpNoDelay(NULL,fd); if (server.tcpkeepalive) anetKeepAlive(NULL,fd,server.tcpkeepalive); if (aeCreateFileEvent(server.el,fd,AE_READABLE, readQueryFromClient, c) == AE_ERR) { close(fd); zfree(c); return NULL; } } selectDb(c,0); c->id = server.next_client_id++; c->fd = fd; c->name = NULL; c->bufpos = 0; c->querybuf = sdsempty(); 初始化是0 readQueryFromClient(networking.c): void readQueryFromClient(aeEventLoop *el, int fd, void *privdata, int mask) { redisClient *c = (redisClient*) privdata; int nread, readlen; size_t qblen; REDIS_NOTUSED(el); REDIS_NOTUSED(mask); server.current_client = c; readlen = REDIS_IOBUF_LEN; /* If this is a multi bulk request, and we are processing a bulk reply * that is large enough, try to maximize the probability that the query * buffer contains exactly the SDS string representing the object, even * at the risk of requiring more read(2) calls. This way the function * processMultiBulkBuffer() can avoid copying buffers to create the * Redis Object representing the argument. */ if (c->reqtype == REDIS_REQ_MULTIBULK && c->multibulklen && c->bulklen != -1 && c->bulklen >= REDIS_MBULK_BIG_ARG) { int remaining = (unsigned)(c->bulklen+2)-sdslen(c->querybuf); if (remaining < readlen) readlen = remaining; } qblen = sdslen(c->querybuf); if (c->querybuf_peak < qblen) c->querybuf_peak = qblen; c->querybuf = sdsMakeRoomFor(c->querybuf, readlen); 在这里会被扩大由此可见c->querybuf在连接第一次读取命令后的大小就会被分配至少1024*32,所以回过头再去看resize的清理逻辑就明显存在问题,每个被使用到的query buffer的大小至少就是1024*32,但是清理的时候判断条件是>1024,也就是说,所有的idle>2的被使用过的连接都会被resize掉,下次接收到请求的时候再重新分配到1024*32,这个其实是没有必要的,在访问比较频繁的群集,内存会被频繁得回收重分配,所以我们尝试将清理的判断条件改造为如下,就可以避免大部分没有必要的resize操作:
if (((querybuf_size > REDIS_MBULK_BIG_ARG) && (querybuf_size/(c->querybuf_peak+1)) > 2) || (querybuf_size > 1024*32 && idletime > 2)) { /* Only resize the query buffer if it is actually wasting space. */ if (sdsavail(c->querybuf) > 1024*32) { c->querybuf = sdsRemoveFreeSpace(c->querybuf); } }这个改造的副作用是内存的开销,按照一个实例5k连接计算,5000*1024*32=160M,这点内存消耗对于上百G内存的服务器完全可以接受。
【问题重现】
在使用修改过源码的Redis server后,问题仍然重现了,客户端还是会报同类型的错误,且报错的时候,服务器内存依然会出现抖动。抓取内存堆栈信息如下:
Thu Jun 14 21:56:54 CST 2018
#3 0x0000003729ee893d in clone () from /lib64/libc.so.6
Thread 1 (Thread 0x7f2dc108d720 (LWP 27851)):
#0 0x0000003729ee5400 in madvise () from /lib64/libc.so.6
#1 0x0000000000493a1e in je_pages_purge ()
#2 0x000000000048cf40 in arena_purge ()
#3 0x00000000004a7dad in je_tcache_bin_flush_large ()
#4 0x00000000004a85e9 in je_tcache_event_hard ()
#5 0x000000000042c0b5 in decrRefCount ()
#6 0x000000000042744d in resetClient ()
#7 0x000000000042963b in processInputBuffer ()
#8 0x0000000000429762 in readQueryFromClient ()
#9 0x000000000041847c in aeProcessEvents ()
#10 0x000000000041873b in aeMain ()
#11 0x0000000000420fce in main ()
Thu Jun 14 21:56:54 CST 2018
Thread 1 (Thread 0x7f2dc108d720 (LWP 27851)):
#0 0x0000003729ee5400 in madvise () from /lib64/libc.so.6
#1 0x0000000000493a1e in je_pages_purge ()
#2 0x000000000048cf40 in arena_purge ()
#3 0x00000000004a7dad in je_tcache_bin_flush_large ()
#4 0x00000000004a85e9 in je_tcache_event_hard ()
#5 0x000000000042c0b5 in decrRefCount ()
#6 0x000000000042744d in resetClient ()
#7 0x000000000042963b in processInputBuffer ()
#8 0x0000000000429762 in readQueryFromClient ()
#9 0x000000000041847c in aeProcessEvents ()
#10 0x000000000041873b in aeMain ()
#11 0x0000000000420fce in main ()
显然,Querybuffer被频繁resize的问题已经得到了优化,但是还是会出现客户端报错。这就又陷入了僵局。难道还有其他因素导致query buffer resize变慢?我们再次抓取pstack。但这时,jemalloc引起了我们的注意。此时回想Redis的内存分配机制,Redis为避免libc内存不被释放导致大量内存碎片的问题,默认使用的是jemalloc用作内存分配管理,这次报错的堆栈信息中都是je_pages_purge () redis在调用jemalloc回收脏页。我们看下jemalloc做了些什么:
arena_purge(arena.c) static void arena_purge(arena_t *arena, bool all) { arena_chunk_t *chunk; size_t npurgatory; if (config_debug) { size_t ndirty = 0; arena_chunk_dirty_iter(&arena->chunks_dirty, NULL, chunks_dirty_iter_cb, (void *)&ndirty); assert(ndirty == arena->ndirty); } assert(arena->ndirty > arena->npurgatory || all); assert((arena->nactive opt_lg_dirty_mult) < (arena->ndirty - arena->npurgatory) || all); if (config_stats) arena->stats.npurge++; npurgatory = arena_compute_npurgatory(arena, all); arena->npurgatory += npurgatory; while (npurgatory > 0) { size_t npurgeable, npurged, nunpurged; /* Get next chunk with dirty pages. */ chunk = arena_chunk_dirty_first(&arena->chunks_dirty); if (chunk == NULL) { arena->npurgatory -= npurgatory; return; } npurgeable = chunk->ndirty; assert(npurgeable != 0); if (npurgeable > npurgatory && chunk->nruns_adjac == 0) { arena->npurgatory += npurgeable - npurgatory; npurgatory = npurgeable; } arena->npurgatory -= npurgeable; npurgatory -= npurgeable; npurged = arena_chunk_purge(arena, chunk, all); nunpurged = npurgeable - npurged; arena->npurgatory += nunpurged; npurgatory += nunpurged; } }Jemalloc每次回收都会判断所有实际应该清理的chunck并对清理做count,这个操作对于高响应要求的系统是很奢侈的,所以我们考虑通过升级jemalloc的版本来优化purge的性能。Redis 4.0版本发布后,性能有很大的改进,并可以通过命令回收内存,我们线上也正准备进行升级,跟随4.0发布的jemalloc版本为4.1,jemalloc的版本使用的在jemalloc的4.0之后版本的arena_purge()做了很多优化,去掉了计数器的调用,简化了很多判断逻辑,增加了arena_stash_dirty()方法合并了之前的计算和判断逻辑,增加了purge_runs_sentinel,用保持脏块在每个arena LRU中的方式替代之前的保持脏块在arena树的dirty-run-containing chunck中的方式,大幅度减少了脏块purge的体积,并且在内存回收过程中不再移动内存块。代码如下:
arena_purge(arena.c) static void arena_purge(arena_t *arena, bool all) { chunk_hooks_t chunk_hooks = chunk_hooks_get(arena); size_t npurge, npurgeable, npurged; arena_runs_dirty_link_t purge_runs_sentinel; extent_node_t purge_chunks_sentinel; arena->purging = true; /* * Calls to arena_dirty_count() are disabled even for debug builds * because overhead grows nonlinearly as memory usage increases. */ if (false && config_debug) { size_t ndirty = arena_dirty_count(arena); assert(ndirty == arena->ndirty); } assert((arena->nactive arena->lg_dirty_mult) < arena->ndirty || all); if (config_stats) arena->stats.npurge++; npurge = arena_compute_npurge(arena, all); qr_new(&purge_runs_sentinel, rd_link); extent_node_dirty_linkage_init(&purge_chunks_sentinel); npurgeable = arena_stash_dirty(arena, &chunk_hooks, all, npurge, &purge_runs_sentinel, &purge_chunks_sentinel); assert(npurgeable >= npurge); npurged = arena_purge_stashed(arena, &chunk_hooks, &purge_runs_sentinel, &purge_chunks_sentinel); assert(npurged == npurgeable); arena_unstash_purged(arena, &chunk_hooks, &purge_runs_sentinel, &purge_chunks_sentinel); arena->purging = false; }【解决问题】
实际上我们有多个选项。可以使用Google的tcmalloc来代替jemalloc,可以升级jemalloc的版本等等。我们根据上面的分析,尝试通过升级jemalloc版本,实际操作为升级Redis版本来解决。我们将Redis的版本升级到4.0.9之后观察,线上客户端连接超时这个棘手的问题得到了解决。
【问题总结】
Redis在生产环境中因其支持高并发,响应快,易操作被广泛使用,对于运维人员而言,其响应时间的要求带来了各种各样的问题,Redis的连接超时问题是其中比较典型的一种,从发现问题,客户端连接超时,到通过抓取客户端与服务端的网络包,内存堆栈定位问题,也被其中一些假象所迷惑,最终通过升级jemalloc(Redis)的版本解决问题,这次最值得总结和借鉴的是整个分析的思路。
总结
以上就是这篇文章的全部内容了,希望本文的内容对大家的学习或者工作具有一定的参考学习价值,如果有疑问大家可以留言交流,谢谢大家对的支持。